PEP 836: JIT Go Brrr: The Path to a Supported JIT Compiler for CPython

11 hours ago

A PEP for a roadmap to an officially supported JIT in CPython has been drafted, responding to the Steering Council's pause on JIT development.
The JIT must meet performance, compatibility, tooling, platform, distribution, security, and maintenance goals, with a 20% geometric mean performance improvement target on pyperformance for JIT + free-threaded builds by Python 3.17's first beta.
Discussions include potential fallback to trace-recorder if method-frontend fails, options like attracting expertise or building in-house, and concerns over experimentation slowing development.
Questions were raised about the C-API's role in JIT performance, type hints' limited utility, and learning from existing JIT projects like Numba.
Memory consumption goals for the JIT are not clearly defined, though there is acknowledgment of the need for an upper limit.
Debate exists over supporting 32-bit Windows, with arguments for deprioritizing it due to declining usage.
The PEP suggests aligning the JIT's optimizer with Single Static Assignment (SSA) form properties, potentially influencing bytecode changes, though this is considered out of scope for the PEP's main goals.

Hasty Briefsbeta